Text-to-Speech
MLX
Safetensors
Chinese
English
audiodit
tts
speech
longcat
diffusion
quantized
int8
apple-silicon
Instructions to use appautomaton/longcat-audiodit-3.5b-8bit-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use appautomaton/longcat-audiodit-3.5b-8bit-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir longcat-audiodit-3.5b-8bit-mlx appautomaton/longcat-audiodit-3.5b-8bit-mlx
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
LongCat AudioDiT 3.5B โ MLX 8-bit
This repository contains a self-contained MLX-native int8 conversion of LongCat AudioDiT 3.5B for local text-to-speech on Apple Silicon.
It is intended for local speech generation with
mlx-speech, without a PyTorch
runtime at inference time.
Model Details
- Developed by: AppAutomaton
- Shared by: AppAutomaton on Hugging Face
- Upstream model:
meituan-longcat/LongCat-AudioDiT-3.5B - Task: text-to-speech and voice cloning
- Runtime: MLX on Apple Silicon
- Precision: int8 quantized weights with bundled tokenizer
Bundle Contents
This bundle is self-contained and includes:
config.jsonmodel.safetensors- tokenizer files (
tokenizer.json,tokenizer_config.json,special_tokens_map.json)
How to Get Started
Basic generation:
python scripts/generate/longcat_audiodit.py \
--text "Hello from LongCat AudioDiT." \
--output-audio outputs/longcat.wav
Voice cloning:
python scripts/generate/longcat_audiodit.py \
--text "Hello from LongCat AudioDiT." \
--prompt-text "Original speaker text." \
--prompt-audio /path/to/prompt.wav \
--output-audio outputs/longcat_clone.wav \
--guidance-method apg
Minimal Python usage:
from pathlib import Path
from mlx_speech.generation.longcat_audiodit import generate_longcat_audiodit
result = generate_longcat_audiodit(
text="Hello from LongCat AudioDiT.",
output_audio="outputs/longcat.wav",
)
Notes
- This repo contains the quantized MLX runtime artifact only.
- The conversion preserves the LongCat AudioDiT diffusion transformer and bundled VAE for waveform decode.
- Voice cloning uses
--guidance-method apg(Adaptive Projected Guidance) orcfg(Classifier-Free Guidance, default).--guidance-strengthcontrols speaker adherence (default: 4.0). - The current bundle is intended for local MLX runtime use and parity validation.
Links
- Source code: mlx-speech
- More examples: AppAutomaton
License
MIT License โ following the upstream license published with
meituan-longcat/LongCat-AudioDiT-3.5B.
- Downloads last month
- 8
Model size
1B params
Tensor type
F32
ยท
U32 ยท
Hardware compatibility
Log In to add your hardware
Quantized
Model tree for appautomaton/longcat-audiodit-3.5b-8bit-mlx
Base model
meituan-longcat/LongCat-AudioDiT-3.5B